Quantized Lite Convolutional Neural Network Hardware Accelerator Design with FPGA for Face Direction Recognition

Abstract

In this study, a quantized lite convolutional neural network (CNN) is applied to accelerate the computations of face direction recognition with Xilinx ZedBoard FPGA platform. The utilized 8-layer lite CNN includes three convolution layers, one max-pooling layer, two average-pooling layers, two fully connected layers. Firstly, the weighting parameters and deviation values of each layer in the CNN are extracted by the training process with software, and the then quantization precision by 8-bit integer (INT8) is used as the inference calculations. By the quantized lite CNN model through the hardware acceleration, the facial direction is correctly detected to achieve fast recognition. Comparison with the software-based implementation in personal computer, the speed-up ratio by the FPGA-based acceleration is about 1.44 times with 0.289 seconds inference time.